Computerized Medical Imaging and Graphics最新文献

{"title":"Integrating AI into clinical practice: Human-centered design requirements for next-generation sequencing workflows","authors":"Markus Plass ,&nbsp;Andreas Holzinger ,&nbsp;Robert Reihs ,&nbsp;Heimo Müller","doi":"10.1016/j.compmedimag.2026.102758","DOIUrl":"10.1016/j.compmedimag.2026.102758","url":null,"abstract":"<div><div>Today, the integration of next-generation sequencing (NGS) into clinical genomics is increasingly AI-driven, with artificial intelligence (AI) underpinning every stage from data processing to decision support. While NGS enables rapid and scalable analysis of complex genetic information—paving the way for precision diagnostics and stratified treatment—the transition from potential to practice is hindered by fragmented workflows, limited usability, and non-standardized data interfaces. This paper introduces a design-oriented framework for embedding AI-powered NGS workflows into clinical decision support systems (CDSS), currently focused on genetic screening and tumor testing, but with components extendable to pathogen detection scenarios. DUXU (Design, User eXperience, Usability) is presented as a conceptual and methodological framework rather than a concrete implementation; its realization is intentionally flexible and must be adapted to the requirements, constraints, and objectives of specific clinical use cases. Future work will adapt data requirements (e.g., taxonomic classification instead of variant calling), functional workflows (e.g., microbial genome assembly), and stakeholder roles (e.g., microbiologists in antimicrobial stewardship). Grounded in real-world clinical environments and aligned with standards such as FHIR and GA4GH, we highlight the central role of AI in multimodal data interpretation, patient-specific visualization, and transparent, explainable decision-making. Anchored in DUXU principles, our approach addresses the socio-technical demands of clinical genomics and proposes actionable design requirements for interoperable, role-specific interfaces. This work advances the development of explainable, interpretable, trustworthy, and operationally embedded AI-based NGS systems into clinical practice.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"130 ","pages":"Article 102758"},"PeriodicalIF":4.9,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147596053","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Agent-MIRA: AI-orchestrated medical imaging agent for PET image retrieval and assistance","authors":"Rajat Vashistha ,&nbsp;Sandra Brosda ,&nbsp;Lauren G. Aoude ,&nbsp;Jessica Ng ,&nbsp;Parveen Kundu ,&nbsp;Andrew P. Barbour ,&nbsp;Viktor Vegh","doi":"10.1016/j.compmedimag.2026.102725","DOIUrl":"10.1016/j.compmedimag.2026.102725","url":null,"abstract":"<div><div>Reporting on medical images can be time-consuming, especially in high-volume clinical settings. AI-agents designated to specific medical imaging tasks can potentially lead to improvements in clinical workflows. We present a prototype AI-agent to support clinical decision making by retrieving medical images which best match the patient images based on similarity and provide uncertainty estimation. The framework requires clinical metadata and PET images with lesion segmentation. A new patient's PET scan is processed by converting it to a feature vector representative of the image, which then enables the retrieval of the nearest feature vector neighbors by querying a database. Comparison between the radiomics and finetuned DINOv2 features was performed. Conditional uncertainty, an estimation based on feature significance, is calculated to state the level of confidence in similarity between patients. The AI agent, using DINOv2-derived features, retrieves a consistent set of patient cases that are phenotypically similar to the new patient. Each retrieved case is accompanied by clinical metadata, including cancer type, treatment history, and survival outcome. It also provides an estimate of the uncertainty in the matches and attention-based visualization to interpret the DINOv2 features. It is validated using eight independent patient test cases with benchmarking via clinical scoring to establish the level of support achieved for AI-agent orchestrated clinical decision-making. Scoring by the clinicians showed good correlation between the new patient and the retrieved database images with respect to the low uncertainty matches. We also integrated image-based retrieval with an entirely parallel text-embedding index of external clinical trials, thereby coupling case-based reasoning with evidence-based medicine in a single query interface using large language model.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"129 ","pages":"Article 102725"},"PeriodicalIF":4.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175371","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Segmentation-aware Generative Reinforcement Network (GRN) for tissue layer segmentation in 3-D ultrasound images for chronic low-back pain (cLBP) assessment","authors":"Zixue Zeng ,&nbsp;Xiaoyan Zhao ,&nbsp;Matthew Cartier ,&nbsp;Tong Yu ,&nbsp;Jing Wang ,&nbsp;Xin Meng ,&nbsp;Zhiyu Sheng ,&nbsp;Maryam Satarpour ,&nbsp;John M. Cormack ,&nbsp;Allison Bean ,&nbsp;Ryan Nussbaum ,&nbsp;Maya Maurer ,&nbsp;Emily Landis-Walkenhorst ,&nbsp;Dinesh Kumbhare ,&nbsp;Kang Kim ,&nbsp;Ajay D. Wasan ,&nbsp;Jiantao Pu","doi":"10.1016/j.compmedimag.2026.102736","DOIUrl":"10.1016/j.compmedimag.2026.102736","url":null,"abstract":"<div><div>Layer-wise segmentation of three-dimensional (3D) ultrasound for chronic lower back pain (cLBP) requires a large amount of labeled images. To mitigate this burden, we propose a Generative Reinforcement Network (GRN) that integrates a generative adversarial network (GAN) framework with a segmentation model. The generator is coupled to a segmentor via segmentation-aware feedback and regularized by a discriminator. At each iteration, the segmentation loss is back-propagated into the generator to produce easy-to-learn reconstructions that directly reduce downstream segmentation error (reinforcement augmentation, RAug), while adversarial feedback from the discriminator (PatchGAN) encourages realistic reconstructions. We also introduce segmentation-guided enhancement (SGE), where the pre-trained generator enhances input images at inference to improve segmentation. GRN has two variants: GRN-SEL, which uses RAug only, and GRN-SSL, which additionally applies interpolation-consistency training (ICT) on unlabeled data by interpolating generator-reconstructed pairs and enforcing prediction consistency. We evaluate GRN primarily on a fully annotated lumbar back ultrasound dataset (MUSCLE). Two public benchmark datasets (skin lesion, Kvasir) were also used to demonstrate its generalizability. On the MUSCLE dataset, GRN-SEL with SGE reduces labeling efforts by up to 70% while improving the Dice Similarity Coefficient (DSC) by 1.98% compared to the models trained on fully labeled datasets. Across all three datasets and label fractions, GRN consistently outperforms state-of-the-art semi-supervised methods. These results suggest the effectiveness of the GRN framework in optimizing segmentation performance with significantly less labeled data. The source code is publicly available at <span><span>https://github.com/Francisdadada/GRN</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"129 ","pages":"Article 102736"},"PeriodicalIF":4.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147357233","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Segmentation-guided multi-modal brain tumors survival prediction model using pseudo-labeling approach","authors":"Ruiquan Ge ,&nbsp;Qingsong Wang ,&nbsp;Xin Lin ,&nbsp;Xinyang Li ,&nbsp;Changmiao Wang ,&nbsp;Zhipeng Wang ,&nbsp;Yuqing Peng ,&nbsp;Xiang Wan ,&nbsp;Ahmed Elazab","doi":"10.1016/j.compmedimag.2026.102724","DOIUrl":"10.1016/j.compmedimag.2026.102724","url":null,"abstract":"<div><div>Predicting a patient’s survival period with a brain tumor can enable doctors to more efficiently assess the efficacy of various treatment modalities and adjust clinical treatment plans under the actual circumstances. This approach aims to enhance patients’ prognosis and survival period. In recent years, there have been significant advancements in predicting the survival period of brain tumors using deep learning techniques. However, most of these methods require the use of two or more independent networks for subsequent knowledge sharing, often overlooking population information. Moreover, analyzing censored data presents an additional challenge, as incomplete information can make it difficult to apply standard analytical techniques and often leads to suboptimal performance in predictive models. To address these challenges, we propose an end-to-end multi-model pseudo-label approach for brain tumor survival prediction, designed to accurately estimate patient survival periods. First, our framework integrates patient population information for optimal use of population information. Next, to enhance data correction, we introduce a method for generating class labels, effectively enlarging the original sample size and improving data utilization efficiency. Additionally, we collected and supplemented survival labels from the BraTS 2021 dataset, which now serves as a new dataset for both brain tumor segmentation and survival prediction. Experimental results demonstrate that our model can enhance the precision of predicting the survival rate of brain tumor patients and exhibit superior generalization. The source code is available at: <span><span>https://github.com/linciel/Brain-Tumor-Survival-Prediction</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"129 ","pages":"Article 102724"},"PeriodicalIF":4.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175369","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"CR-GLoCo: Cross-Resolution Learning via Global-Local Context Consistency for semi-supervised 3D medical segmentation","authors":"Peng Liu,&nbsp;Guoyan Zheng","doi":"10.1016/j.compmedimag.2026.102742","DOIUrl":"10.1016/j.compmedimag.2026.102742","url":null,"abstract":"<div><div>Deep learning-based methods have achieved remarkable accuracy in segmenting three-dimensional (3D) medical images but they typically rely on large amounts of expert annotated datasets for training. Annotating such datasets, particularly at the voxel-level, is laborious and costly. To alleviate such annotation burden, semi-supervised learning (SSL) is introduced, which can leverage a small set of labeled data together with abundant unlabeled data for improved performance. Limited by GPU memory, patch- or slice-based strategies are often adopted to train such a SSL model, which may result in a loss of global anatomical context. Additionally, patch/slice-based losses often make the trained SSL models sensitive to contextual changes and hinder generalization under scarce labels. To this end, we propose CR-GLoCo, a cross-resolution learning framework that enforces Global-Local Context Consistency for semi-supervised 3D medical image segmentation. CR-GLoCo couples a low-resolution global branch with a high-resolution local branch and enforces prediction consistency on their overlapping field of view. In particular, the global branch provides holistic anatomical priors while the local branch preserves fine boundaries. We further design a mutual pseudo-supervision with confidence filtering to transfer reliable cues across resolutions and adopt an overlap-based sampling strategy to leverage changing contexts during training. We conduct comprehensive experiments on three typical yet challenging datasets to evaluate the performance of the proposed CR-GLoCo framework. Experimental results demonstrate that our proposed method achieves superior performance than other state-of-the-art SSL methods. Our code is available at <span><span>https://github.com/luckieucas/CR-GLoCo</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"129 ","pages":"Article 102742"},"PeriodicalIF":4.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147357271","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"VFA-Net3D: A zero-shot vascular flow-guided 3D network for brain vessel segmentation in acute ischemic stroke","authors":"Asim Zaman ,&nbsp;Mazen M. Yassin ,&nbsp;Rashid Khan ,&nbsp;Faizan Ahmad ,&nbsp;Guangtao Huang ,&nbsp;Yongkang Shi ,&nbsp;Ziran Chen ,&nbsp;Yan Kang","doi":"10.1016/j.compmedimag.2026.102712","DOIUrl":"10.1016/j.compmedimag.2026.102712","url":null,"abstract":"<div><div>Timely intervention in acute ischemic stroke (AIS) is critical, with approximately 1.9 million neurons lost per minute. Clinical time-of-flight magnetic resonance angiography (TOF-MRA) protocols, designed for rapid acquisition within a few minutes, introduce substantial domain shifts compared to high-resolution research datasets. These include reduced resolution, a lower signal-to-noise ratio, partial-volume effects, and motion artifacts, which are compounded by stroke-specific vascular abnormalities. Conventional segmentation models often fail under such conditions due to limited robustness to domain variability. We propose Vascular Flow-Attention Network (VFA-Net), a fully 3D deep neural network designed for zero-shot vessel segmentation in AIS, enabling knowledge transfer from annotated healthy TOF-MRA scans without retraining on pathological or low-quality clinical data. The architecture integrates five novel modules: (1) Flow-Pattern Attention for vascular continuity; (2) Multi-Scale Context Aggregation using dilated attention; (3) Vascular Flow Feature Refinement for adaptive attention enhancement; (4) Boundary-Guided Skip for precise boundary delineation; and (5) Flow-Refined Up-sampling to recover fine vessel details. The proposed model, trained exclusively on healthy TOF-MRA scans from four vendors (1.5 T and 3 T), was evaluated across three experimental configurations and significantly outperformed state-of-the-art models. By explicitly encoding vascular domain knowledge (continuity, boundaries, and flow topology) into its architecture, VFA-Net3D functions as a knowledge-guided system, enabling robust zero-shot generalization across clinical domains. VFA-Net3D presents a robust and clinically deployable solution for AIS vessel segmentation, supporting faster and more accurate diagnosis and treatment planning.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"129 ","pages":"Article 102712"},"PeriodicalIF":4.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146147488","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"NEC-Net: A hybrid transformer for necrotizing enterocolitis diagnosis by lesion segmentation on abdominal X-rays","authors":"Guo Jing ,&nbsp;Li Yunjiao ,&nbsp;Wenhai Fang ,&nbsp;Long Xiongbai ,&nbsp;Zhong Wei ,&nbsp;He Qiuming ,&nbsp;Ren Hongliang","doi":"10.1016/j.compmedimag.2026.102731","DOIUrl":"10.1016/j.compmedimag.2026.102731","url":null,"abstract":"<div><div>Accurate lesion segmentation in medical imaging is essential for computer-aided diagnosis and clinical decision-making. In particular, necrotizing enterocolitis (NEC), a life-threatening gastrointestinal disease in neonates, requires precise localization of lesion areas from abdominal X-rays to support timely intervention. However, NEC lesion segmentation is highly challenging due to local feature similarity, indistinct boundaries, and overlapping tissues in radiographic images. To address these issues, we propose NEC-Net, a hybrid CNN-Transformer architecture specifically designed for fine-grained NEC lesion segmentation: a Series–Parallel Large Kernel Attention module to enhance the representation of elongated structures, a Multi-layer Edge-enhanced Perception Module for improved boundary sensitivity, and a Short-range Concatenation Decoding structure to better handle small lesions. Extensive experiments on a clinical dataset comprising 752 annotated NEC cases demonstrate the superiority of NEC-Net, achieving 77.92% mIoU, 86.14% mean Dice coefficient and 35.31 HD95. The model, along with associated tools, is available at <span><span>https://github.com/priscillaLee0830/NEC-application</span><svg><path></path></svg></span>, offering practical value for NEC radiological alert and clinical deployment.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"129 ","pages":"Article 102731"},"PeriodicalIF":4.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147312000","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Scale-consistent 3D reconstruction in monocular colonoscopy via camera-intrinsics-guided learning","authors":"Yeqi Liu ,&nbsp;Deping Yu ,&nbsp;Yijiang Xiao ,&nbsp;Tian He ,&nbsp;Ling Liu","doi":"10.1016/j.compmedimag.2026.102734","DOIUrl":"10.1016/j.compmedimag.2026.102734","url":null,"abstract":"<div><div>Reliable spatial understanding of the colon lumen is crucial for geometry-aware analysis and quality assessment in colonoscopy. However, monocular colonoscopes are not equipped with depth sensors, and existing methods struggle to recover depth and camera motion with a stable, consistent scale. To address this challenge, we propose a novel framework for 3D reconstruction in monocular colonoscopy, aiming to recover scale-consistent depth and camera motion in an intrinsics-defined scale space. The framework adopts a two-step training strategy. Step 1, a depth estimation network is pretrained on synthetic colonoscopy data derived from CT-based colon models. We introduce an Adaptive Depth Adjustment module to better match endoscopic depth characteristics and to produce more stable depth predictions. The pretrained model is then applied to real images to predict depth under the synthetic-camera setting, and an intrinsics-guided scale alignment module converts these predictions into pseudo-depth labels that are interpretable in the target intrinsics-defined scale space. Step 2, these pseudo-labels guide joint depth–pose learning on real sequences with a geometry-aware 3D loss to enforce multi-view consistency and suppress scale drift. During inference, a globally consistent optimization strategy mitigates long-sequence drift. Experiments on C3VD show state-of-the-art depth and pose accuracy, reducing AbsRel by 16.1% and trajectory APE by 86.7% over the strongest baseline. Evaluations on clinical colonoscopy videos show that the reconstructed lumen geometry exhibits anatomically plausible diameter ranges, indicating that the recovered scale is stable and physically interpretable under known camera intrinsics. Overall, the proposed framework provides a sensor-free way to obtain scale-consistent depth and motion under a known imaging model, enabling more reliable geometry-aware reconstruction and downstream analysis in monocular colonoscopy.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"129 ","pages":"Article 102734"},"PeriodicalIF":4.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147318952","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Transformer-based architectures in MRI brain tumor segmentation: A review","authors":"Chengcheng Jin ,&nbsp;Nor Safira Elaina Mohd Noor ,&nbsp;Theam Foo Ng ,&nbsp;Mohd Shahrimie Mohd Asaari ,&nbsp;Haidi Ibrahim","doi":"10.1016/j.compmedimag.2026.102729","DOIUrl":"10.1016/j.compmedimag.2026.102729","url":null,"abstract":"<div><div>Transformers have been actively utilized in the deep learning field recently. Vision Transformer (ViT), as one of its important applications in the computer vision field, exhibits significant promise for automatic glioma MRI image segmentation. As the Transformer has a large efficient receptive field, it can focus on the tumor range as well as surrounding tissues and organs. Hence, many variant models derived from Transformer architecture have been developed for medical image segmentation. Effective model architectures can significantly enhance segmentation performance, and Swin Transformer is one typical structural optimization. Conversely, since both Transformer and U-Net have significant applications in medical image segmentation and enable seamless integration, their combination has become the predominant architectural design strategy. Furthermore, effective self-attention mechanisms have a strong ability to capture features, and the size and position of the patch also significantly influence the efficacy of ViT. In general, this paper analyzes the applications of the Transformer variant algorithms from model architecture design, efficient self-attention mechanism, and patch acquisition strategy. This paper concentrates on the applications of Transformers for glioma MRI segmentation, to give researchers in the area a reference and method comparisons.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"129 ","pages":"Article 102729"},"PeriodicalIF":4.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147272693","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Mamba‑MFNet: A hierarchical supervised network based on the fusion of axial and cross-modal attention for breast DCE‑MRI tumor segmentation","authors":"Yuxin Gao ,&nbsp;Halidanmu Abudukelimu ,&nbsp;Guzhanuer Abudouxikuer ,&nbsp;Mayilamu Musideke ,&nbsp;Mengfei Wang ,&nbsp;Mireguli Aizizi ,&nbsp;Abudukelimu Abulizi ,&nbsp;Mayila Abudukelimu","doi":"10.1016/j.compmedimag.2026.102732","DOIUrl":"10.1016/j.compmedimag.2026.102732","url":null,"abstract":"<div><div>Early detection and accurate segmentation of breast cancer are critical for improving cure rates and optimizing treatment strategies. However, existing 3D segmentation methods face significant limitations in global context modeling and efficient multimodal feature fusion. To address these challenges, this study proposes Mamba‑MFNet, a 3D breast tumor segmentation framework based on a state-space model, which introduces three key innovations. First, a cross-modal attention mechanism dynamically fuses six DCE-MRI temporal modalities with one breast anatomical mask and adaptively adjusts the contribution of each channel using a learnable similarity matrix and weight parameters. Second, a three-dimensional axial attention module computes multi-head attention along the coronal, sagittal, and transverse planes, and reorganizing features through convolutions to model global dependencies with linear complexity, effectively filling in a blind spot of the local convolution receptive field in capturing long-range contex. Finally, Hierarchical loss strategy is adopted in the high-resolution layer, combining multi-scale deep supervision and boundary-aware loss, introducing focal loss in the middle layer, and focusing on global semantics in the low-resolution layer, which significantly improves the detection accuracy of tiny lesions and blurred boundaries. Experimental results on the breast DCE-MRI dataset validate the effectiveness of the proposed method: compared with baseline UMamba, Mamba-MFNet improved the Dice coefficient, Precision, Recall, and IoU by 5.07%, 2.84%, 4.75%, and 3.39%, respectively. Additionally, nine clinical experts commended highly the model’s accuracy, boundary clarity, and clinical decision support capability in a subjective evaluation, further demonstrating the reliability and usefulness of Mamba-MFNet in real-world diagnostic and treatment scenarios.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"129 ","pages":"Article 102732"},"PeriodicalIF":4.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147327997","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}